Hierarchical probabilistic task recognition based on spatial memory for care support

We propose a method for recognizing tasks performed by care workers. Time-series sample data for each feature amount during tasks are defined as a task history, which is used as a basis for creating spatiotemporal task information in reference to Niitsuma's spatial memory. We perform simulated care tasks in an environment that recreates an actual care site, and measure time-series data of worker feature amounts by a motion capture system. We divide this data into learning and evaluation data, and verify the recognition accuracy. Recognition accuracy for seven defined elemental tasks is close to 80% on average, demonstrating the effectiveness of this method.